Pith. sign in

REVIEW 4 major objections 4 minor 54 references

By backpropagating a constraint loss through the reverse denoising rollout, this paper shows that a diffusion model can be fine-tuned to satisfy hard feasibility constraints under its own sampling dynamics, achieving near-zero violations wh

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:11 UTC pith:RPN7ULVO

load-bearing objection Rollout fine-tuning for constrained diffusion looks promising empirically, but the proof of the main convergence proposition is not valid, so the theory should be treated as heuristic until fixed. the 4 major comments →

arxiv 2607.14398 v1 pith:RPN7ULVO submitted 2026-07-15 cs.LG cs.AI

Integration Matters: Rollout-Based Training for Constrained Diffusion Models

Xiaoxuan Liang , Saeid Naderiparizi , Berend Zwartsenberg , Frank Wood This is my paper
classification cs.LG cs.AI
keywords constrained diffusion modelsrollout-based fine-tuningconstraint satisfactiondenoising trajectory alignmentlearned guidance scalingtrajectory predictiondiffusion model fine-tuningfeasibility constraints
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Constrained diffusion models need to generate samples that satisfy hard feasibility constraints (e.g., on-road, collision-free trajectories) without drifting from the data distribution. This paper claims that the usual ways of enforcing constraints—penalizing denoising predictions during training or correcting trajectories at sampling time—are misaligned: training sees states from the forward noising process that are near-feasible, while sampling encounters infeasible states along the reverse denoising rollout. The proposed method, IMCD, fine-tunes a pretrained denoiser by differentiating a terminal constraint loss through the full reverse denoising trajectory, so the model learns to correct violations under its own sampling dynamics. If correct, this would give near-zero constraint violations with few sampling steps while preserving sample quality and avoiding the distortion and tuning of sampling-time guidance.

Core claim

The paper establishes that optimizing constraint violations on terminal rollout samples—backpropagating through the fixed noise schedule of the reverse process—aligns training with inference. It shows that with a learned, state-and-time-dependent scaling of a frozen pretrained denoiser's constraint gradient, the model can be fine-tuned to produce near-zero boundary/overlap violations in bouncing-ball generation and near-zero offroad/collision rates in traffic trajectory prediction, while keeping fidelity metrics close to the unconstrained EDM baseline.

What carries the argument

The load-bearing machinery is the rollout objective: a terminal constraint loss L_rollout(φ) = E[ℓ_Ω(x̃_0)] taken over the composed, differentiable denoising map F_φ, with gradients flowing through every step of the reverse trajectory (Theorem 4.3). The constraint-aware denoiser D_φ^Ω combines a frozen guidance gradient ∂₁ℓ_Ω(D(x̃_i; t_i)) from the pretrained denoiser with a learned scaling network γ_φ(x_t; t) = α_φ(x_t; t) · t^{β_φ(x_t; t)}, whose output grows as t→0 to enforce constraints near the data manifold. The EDM denoising loss is retained as a regularizer to preserve data fidelity.

Load-bearing premise

The constraint-gradient direction used to guide the fine-tuned model is computed once with the frozen pretrained denoiser and is assumed to remain a valid, correctly rescaled correction for the evolving fine-tuned model; if the fine-tuned model drifts far from the pretrained one, this guidance can become stale.

What would settle it

During IMCD fine-tuning, compute at matched rollout states the constraint-gradient direction from the frozen pretrained denoiser and from the current fine-tuned denoiser; if their cosine similarity falls markedly while terminal violations stop decreasing (or rise), the frozen-guidance assumption fails. Simpler: ablate the gradient-embedding input at inference; if violations stay near zero, the guidance input is not load-bearing; if they explode, it is—both results change the interpretation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Training-time alignment: the model is exposed to off-manifold states it will actually encounter at inference, so few-step sampling no longer requires strong incremental corrections that distort motion.
  • Sampling-time correction tuning is replaced by learned scaling; there is no grid search over guidance schedules or per-step correction strengths.
  • Constraint satisfaction is achieved without the severe physical distortion (high frame-to-frame jitter and energy deviation) that typifies inference-only correction methods.
  • The approach applies where constraints are given only as violation functions (e.g., drivable meshes) rather than analytic boundaries, extending feasibility enforcement to realistic tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the frozen-denoiser guidance direction remains reliable, the final constraint-aware denoiser could be distilled into a standalone network, eliminating the remaining inference-time gradient computation entirely—a direction the paper itself flags as future work.
  • The learned scaling network's behavior differs across tasks (strong late low-noise corrections in bouncing balls, a mid-trajectory peak in driving), suggesting that correction timing is task-specific; one could test whether a scaling network transfers across related tasks to cut fine-tuning cost.
  • Rollout-based training may reduce exposure bias in unconstrained diffusion generation altogether, since it trains the model under its own sampling dynamics; constraint losses are just the first demonstrated application of this principle.
  • A concrete testable extension: during fine-tuning, monitor the cosine similarity between constraint-gradient directions from the frozen pretrained denoiser and the current fine-tuned denoiser at matched rollout states; if similarity degrades while violations stall or rise, the guidance reference needs periodic refreshment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes IMCD, a fine-tuning method for constrained diffusion models. Given a pretrained denoiser D, it adds lightweight trainable components (LoRA adapters, a gradient embedding, and a learned state- and time-dependent scaling network γϕ) to form a constraint-aware denoiser DΩϕ. The training objective combines the standard EDM denoising loss with a terminal constraint loss evaluated on samples obtained by rolling out the reverse denoising process under the model being fine-tuned; gradients are backpropagated through the fixed discretized trajectory. The paper claims that this rollout-based training aligns training with inference, avoiding the train/sample mismatch of forward-state constraint fine-tuning and the fragility of sampling-time correction. Proposition 4.2 and Theorem 4.3 are offered as formal support for terminal convergence and gradient correctness. Experiments on bouncing balls and INTERACTION driving scenarios show large reductions in violation rates while keeping fidelity close to the unconstrained EDM baseline, and an ablation (Table 4) indicates that rollout-based training is important compared with forward-process training.

Significance. The empirical contribution is substantial and well structured. Table 1 and Table 2 show that IMCD reduces boundary/overlap and offroad/collision violations to near zero while remaining close to EDM on r-ELBO/ADE/FDE, and the FP-versus-rollout ablation in Table 4 supports the central alignment hypothesis. The additional scenarios in Section E.5 and the breadth of baselines make the experiments informative. If the theoretical claims can be repaired or explicitly repositioned as heuristics, the method is likely publishable. However, the current formal analysis is not sound: Proposition 4.2 is mathematically incorrect as stated, and the theory is not connected to the actual parameterization in Equation (13). This is a load-bearing gap because the abstract and contributions advertise formal convergence to the feasible region.

major comments (4)
  1. [§4.1, Prop. 4.2; §B.1] The claimed equality (11) is false as stated. Because γϕ(x_t;t) is a learned function of x_t, the score of pΩ(x_t;t) ∝ p(x_t;t) exp(−γϕ(x_t;t) ℓΩ(D(x_t;t))) contains the additional terms −∇_{x_t}γϕ · ℓΩ and −γϕ (∂D/∂x_t)^T ∂_1ℓΩ, in addition to −γϕ ∂_1ℓΩ. The proof in §B.1 drops the first term by asserting 'treat γ_t as constant' and ignores the second even though Eq. (8) explicitly distinguishes ∂_1ℓΩ from ∇_{x_t}ℓΩ. Consequently the limit (12) is not established. Since the abstract and contributions advertise formal convergence to the feasible region, this is a load-bearing gap; the proposition needs corrected assumptions and additional terms, or the claim must be explicitly downgraded to a heuristic.
  2. [§4.2, Eq. (13); §4.1, Eq. (10)] The score estimator analyzed in Proposition 4.2 is sΩϕ = sϕ − γϕ ∂_1ℓΩ(D). The implemented constraint-aware denoiser is DΩϕ = Dϕ(˜x_i; t_i, g_i) + γϕ g_i with g_i = ∂_1ℓΩ(D(˜x_i); t_i). Under Tweedie's relation these are generally different objects; no derivation shows that the score implied by Eq. (13) equals Eq. (10), and the output-level addition in Eq. (13) has a different sign/scaling from a score subtraction. The formal analysis is therefore disconnected from the network actually trained. Please either connect the two formulations or state explicitly that Proposition 4.2 applies only to an idealized variant.
  3. [§3.2, Table 1, §D.3] The MPGD baseline is evaluated 'without projection' while Table 1 labels the row MPGD. Projection is a core component of MPGD as originally proposed, so the current comparison does not establish that sampling-time correction methods are fragile under few-step sampling; it establishes that this ablated variant is fragile. Either evaluate MPGD with its projection or clearly report the omission in the main table and adjust the Section 3.2 motivation accordingly.
  4. [§4.1, Eq. (8); Algorithm 1] The guidance direction GΩ = ∂_1ℓΩ(D(·)) is always computed with the frozen pretrained denoiser D, while the model being trained is Dϕ. The 'learned rescaling of the exact gradient' interpretation via Eq. (8) is valid only for the Jacobian of D; once Dϕ drifts, GΩ is not the rescaled exact gradient for Dϕ. No experiment or diagnostic measures this drift. This is a correctness risk for the central mechanism; please add a quantitative check (e.g., agreement between GΩ and ∇_{x_t}ℓΩ(Dϕ(x_t)) over training) or discuss the limitation explicitly.
minor comments (4)
  1. [Eq. (12); §B.1] At t→0 the indicator should be applied to x0 (or D(x_t;t)), not to x_t. Also, Eq. (21) inconsistently writes ℓΩ(x_t;t) instead of ℓΩ(D(x_t;t)).
  2. [Theorem 4.3] Theorem 4.3 is essentially the chain rule for composed stochastic maps; calling it an 'alignment theorem' overstates its content. The actual empirical evidence for alignment is the FP-versus-rollout ablation in Table 4. Consider presenting the result as a gradient identity rather than a convergence/alignment theorem.
  3. [Definition 4.1; §D.4, §E.3] The condition βϕ<−1 is stronger than needed for t^β to diverge as t→0 (any β<0 suffices). The relationship between the bounded exponent parameterizations (−4 in bouncing balls, −3 in traffic) and the stated β<-1 condition is not explained; please clarify whether the bound is essential or just a design choice.
  4. [Section 5.2; Table 7] The statement that physical plausibility remains close to EDM is scenario-dependent. In the DR_USA_Roundabout_FT results (Table 7), IMCD has SF2F=1.37 versus EDM=1.23 and SMLV=0.75 versus 0.47. The main text should qualify the claim rather than stating it globally.

Circularity Check

1 steps flagged

Convergence property in Proposition 4.2 is written into the γφ parameterization; the core rollout-based empirical claim remains self-contained.

specific steps
  1. self definitional [Section 4.1, Definition 4.1 and Proposition 4.2; Appendix B.1; Section D.4]
    "Definition 4.1: γϕ(xt;t) = αϕ(xt;t)·t^{βϕ(xt;t)}, where αϕ(xt;t)≥0, βϕ<−1. Proposition 4.2: 'Furthermore, γϕ(xt;t)→∞ as t→0 for any xt with αϕ(xt;t)>0, such that lim_{t→0} pΩ(xt;t) ∝ lim_{t→0} p(xt;t)1_Ω(xt).' Proof B.1: 'since βϕ<−1, t^{βϕ}→∞ as t→0.' D.4: 'This parameterization preserves the desired divergence property as noise level approaches zero.'"

    The advertised convergence of the terminal denoising state toward the feasible region is not derived from the rollout objective or from score analysis; it is inserted directly into the definition of the scaling network. Definition 4.1 forces βϕ<−1, which makes t^{βϕ}→∞ as t→0, and Section D.4 states that the t^{−4} divergence is fixed analytically to 'preserve the desired divergence property.' Proposition 4.2 then restates this imposed condition as a theorem: γ→∞ by definition, so the exponential in Eq. (21) collapses to the indicator, giving Eq. (22). Thus the paper's contribution bullet claiming 'formal analysis ... show that the terminal denoising state converges towards the feasible region' reduces to the parameterization choice rather than being an independent consequence of the metho

full rationale

The empirical core of the paper is not circular: the rollout loss in Eq. (16) directly minimizes the terminal constraint metric, so the reported near-zero boundary/overlap and offroad/collision rates are the result of optimizing that objective, and the fidelity/physical-plausibility metrics are not directly optimized and are evaluated against held-out data. This is standard training-objective evaluation, not a renamed prediction. The guidance function is borrowed from the authors' own MBM++ (Liang et al. [26]) with explicit citation, and the experiments compare against MBM++ and other methods; this is legitimate self-citation and is not load-bearing in the sense of an unverified uniqueness claim. The one genuinely circular element is Proposition 4.2's convergence property: the parameterization βϕ<−1 is chosen precisely to make γϕ diverge, and D.4 admits the t^{−4} divergence is 'fixed analytically,' so the formal convergence result is an artifact of the definition. Separately, the proof in B.1 that Eq. (10) is the score of Eq. (11) is mathematically unsupported because it 'treat[s] γt(xt;t) as constant with respect to xt' while Definition 4.1 defines γϕ as state-dependent; I treat this as a correctness defect, not a circularity, and it does not by itself raise the circularity score. Overall, the main empirical contribution stands independently, but the advertised theoretical convergence analysis is partially self-definitional, giving a score of 4 rather than 0-2.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The method's load-bearing inputs are the frozen pretrained denoiser, the differentiable constraint functions, and the assumption that training and inference share the same noise schedule. The hand-chosen exponent offset is the main free parameter; no new physical entities are introduced.

free parameters (3)
  • exponent offset in βφ parameterization = -4 for bouncing ball (2sigmoid(β_raw)−4); -3 for traffic (2sigmoid(β_raw)−3)
    Hand-chosen per task to set the divergence strength of the scaling network at low noise; the -4 offset is justified by the t^{-4} score/Jacobian scaling, while -3 is a milder choice for trajectory prediction. This is a task-specific design choice, not derived from first principles.
  • division stabilizer ε in κ and gradient projections = 1e-5
    Small constant used to prevent division by zero in Eq (17) and appendix Eq (37); chosen by hand.
  • number of rollout steps N = 50 (bouncing ball), 20 (traffic)
    Matches the inference-time sampler's noise schedule; the paper does not study sensitivity to N for the training rollout.
axioms (6)
  • domain assumption The data distribution is supported on the feasible set: supp(p0) ⊆ Ω.
    Section 2: 'We assume that the data distribution is supported on the feasible set, supp(p0) ⊆ Ω.' This makes the problem well-posed; without it, constraint-satisfying samples could contradict the data distribution.
  • domain assumption The denoiser Jacobian is approximated by a covariance term: ∂D(xt;t)/∂xt ≈ cov[x0|xt]/t².
    Section 4.1: 'Prior work [13,5] shows that the Jacobian scales with a posterior covariance term...' This approximation motivates the functional form of the learned scaling γφ. Not used in the actual training gradient, but underpins the interpretation in Prop 4.2.
  • domain assumption Constraint violation functions ℓΩ are differentiable (or subdifferentiable) at evaluated points.
    The method requires ∂1ℓΩ(D(xt;t);t). For the traffic offroad loss, the indicator 1[c(a)_τ,j ∈ M] in Eq (54) is not differentiable; the paper does not discuss how gradients are obtained across this discontinuity.
  • standard math The composed denoising map Fφ is differentiable in ϕ and the expected-gradient interchange is valid.
    Theorem 4.3 proof (B.2) freely exchanges ∇ϕ and E_{p(x_N)}. This is standard under smoothness and dominated convergence, but the paper states no regularity conditions; for practical neural networks it is usually justified by the reparameterization trick.
  • standard math Tweedie's formula and the EDM forward-process marginal p(xt|x0) = N(x0, t²I) are exact for the chosen noise schedule.
    Used in Eq (5) and the score parameterization; standard diffusion-model background.
  • domain assumption The fixed noise schedule used in training rollouts is the same as the inference-time sampler.
    Section 4.3 and Algorithm 1 use the same discretization; if inference changed steps or schedule, the alignment guarantee would break. The paper keeps them matched in experiments.

pith-pipeline@v1.3.0-alltime-deepseek · 20523 in / 22931 out tokens · 213648 ms · 2026-08-02T02:11:44.247880+00:00 · methodology

0 comments
read the original abstract

Constrained generative models aim to produce samples that satisfy complex feasibility constraints while remaining faithful to the data distribution. Existing constrained generation methods typically enforce constraints either through training-time optimization or sampling-time correction. Training-time optimization approaches optimize on states induced by the training distribution, which can differ substantially from those encountered during sampling. Sampling-time correction methods instead modify the sampling process at inference, introducing distribution shift and requiring expensive tuning, particularly for few-step sampling. We propose a fine-tuning framework that incorporates constraint guidance obtained through online rollout into the training process, which aligns training with sampling by differentiating through the fixed noise schedule used to numerically integrate the denoising process. This exposes the model to violations that arise along the denoising trajectory and aligns diffusion learning with the sampling process. Experiments across multiple tasks show that our method improves constraint satisfaction while maintaining competitive sampling quality compared to prior methods.

Figures

Figures reproduced from arXiv: 2607.14398 by Berend Zwartsenberg, Frank Wood, Saeid Naderiparizi, Xiaoxuan Liang.

Figure 1
Figure 1. Figure 1: Infraction loss of the one-step lookahead [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scaling network output versus noise level [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Scaling network output versus noise level [PITH_FULL_IMAGE:figures/full_fig_p021_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 17 linked inside Pith

  1. [1]

    Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    Brian DO Anderson. Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    1982

  2. [2]

    Physics-informed diffusion models.arXiv preprint arXiv:2403.14404, 2024

    Jan-Hendrik Bastek, WaiChing Sun, and Dennis M Kochmann. Physics-informed diffusion models.arXiv preprint arXiv:2403.14404, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Longformer: The long-document transformer

    Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020

    Pith/arXiv arXiv 2004

  4. [4]

    Dynamical regimes of diffusion models.Nature Communications, 15(1):9957, 2024

    Giulio Biroli, Tony Bonnaire, Valentin De Bortoli, and Marc Mézard. Dynamical regimes of diffusion models.Nature Communications, 15(1):9957, 2024

    2024

  5. [5]

    Tweedie moment projected diffusions for inverse problems.arXiv preprint arXiv:2310.06721, 2023

    Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, and O Deniz Akyildiz. Tweedie moment projected diffusions for inverse problems.arXiv preprint arXiv:2310.06721, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Motion planning diffusion: Learning and planning of robot motions with diffusion models

    Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1916–1923. IEEE, 2023

    1916

  7. [7]

    Training deep nets with sublinear memory cost.arXiv preprint arXiv:1604.06174, 2016

    Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost.arXiv preprint arXiv:1604.06174, 2016

    Pith/arXiv arXiv 2016

  8. [8]

    Projected generative diffusion models for constraint satisfaction.arXiv preprint arXiv:2402.03559, 2024

    Jacob K Christopher, Stephen Baek, and Ferdinando Fioretto. Projected generative diffusion models for constraint satisfaction.arXiv preprint arXiv:2402.03559, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Diffu- sion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffu- sion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    Cfg++: Manifold-constrained classifier free guidance for diffusion models.arXiv preprint arXiv:2406.08070, 2024

    Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models.arXiv preprint arXiv:2406.08070, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  12. [12]

    Adjoint matching: Fine-tuning flow and diffusion generative models with memoryless stochastic optimal control

    Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky TQ Chen. Adjoint matching: Fine-tuning flow and diffusion generative models with memoryless stochastic optimal control. arXiv preprint arXiv:2409.08861, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106(496):1602–1614, 2011

    Bradley Efron. Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106(496):1602–1614, 2011

    2011

  14. [14]

    Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models.Advances in Neural Information Processing Systems, 36:79858–79885, 2023

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models.Advances in Neural Information Processing Systems, 36:79858–79885, 2023

    2023

  15. [15]

    Diffusion models for constrained domains.Transactions on Machine Learning Research, 2023

    Nic Fishman, Leo Klarner, Valentin De Bortoli, Emile Mathieu, and Michael John Hutchinson. Diffusion models for constrained domains.Transactions on Machine Learning Research, 2023

    2023

  16. [16]

    Deep temporal sigmoid belief networks for sequence modeling.Advances in Neural Information Processing Systems, 28, 2015

    Zhe Gan, Chunyuan Li, Ricardo Henao, David E Carlson, and Lawrence Carin. Deep temporal sigmoid belief networks for sequence modeling.Advances in Neural Information Processing Systems, 28, 2015. 10

    2015

  17. [17]

    Flexible diffusion modeling of long videos.Advances in neural information processing systems, 35:27953–27965, 2022

    William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos.Advances in neural information processing systems, 35:27953–27965, 2022

    2022

  18. [18]

    Adjoint sampling: Highly scalable diffusion samplers via adjoint matching.arXiv preprint arXiv:2504.11713, 2025

    Aaron Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Brandon Wood, Daniel Levine, Bin Hu, Brandon Amos, Brian Karrer, et al. Adjoint sampling: Highly scalable diffusion samplers via adjoint matching.arXiv preprint arXiv:2504.11713, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Manifold preserv- ing guided diffusion

    Yutong He, Naoki Murata, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Dongjun Kim, Wei-Hsiang Liao, Yuki Mitsufuji, J Zico Kolter, Ruslan Salakhutdinov, et al. Manifold preserv- ing guided diffusion. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  20. [20]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  21. [21]

    Video diffusion models.Advances in neural information processing systems, 35: 8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models.Advances in neural information processing systems, 35: 8633–8646, 2022

    2022

  22. [22]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    2022

  23. [23]

    Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35: 26565–26577, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35: 26565–26577, 2022

    2022

  24. [24]

    Alleviating exposure bias in diffusion models through sampling with shifted time steps.arXiv preprint arXiv:2305.15583, 2023

    Mingxiao Li, Tingyu Qu, Ruicong Yao, Wei Sun, and Marie-Francine Moens. Alleviating exposure bias in diffusion models through sampling with shifted time steps.arXiv preprint arXiv:2305.15583, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Simultaneous multi-robot motion planning with projected diffusion models.arXiv preprint arXiv:2502.03607, 2025

    Jinhao Liang, Jacob K Christopher, Sven Koenig, and Ferdinando Fioretto. Simultaneous multi-robot motion planning with projected diffusion models.arXiv preprint arXiv:2502.03607, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Improved constrained generation by bridging pretrained generative models.arXiv preprint arXiv:2603.06742, 2026

    Xiaoxuan Liang, Saeid Naderiparizi, Yunpeng Liu, Berend Zwartsenberg, and Frank Wood. Improved constrained generation by bridging pretrained generative models.arXiv preprint arXiv:2603.06742, 2026

    arXiv 2026

  27. [27]

    Critic sequential monte carlo

    Vasileios Lioutas, Jonathan Wilder Lavington, Justice Sefas, Matthew Niedoba, Yunpeng Liu, Berend Zwartsenberg, Setareh Dabiri, Frank Wood, and Adam Scibior. Critic sequential monte carlo. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  28. [28]

    Adjoint schr\" odinger bridge sampler.arXiv preprint arXiv:2506.22565, 2025

    Guan-Horng Liu, Jaemoo Choi, Yongxin Chen, Benjamin Kurt Miller, and Ricky TQ Chen. Adjoint schr\" odinger bridge sampler.arXiv preprint arXiv:2506.22565, 2025

    arXiv 2025

  29. [29]

    Learning diffusion bridges on constrained domains

    Xingchao Liu and Lemeng Wu. Learning diffusion bridges on constrained domains. In international conference on learning representations (ICLR), 2023

    2023

  30. [30]

    Reflected diffusion models

    Aaron Lou and Stefano Ermon. Reflected diffusion models. InInternational Conference on Machine Learning, pages 22675–22701. PMLR, 2023

    2023

  31. [31]

    Amortized rejection sampling in universal probabilistic programming

    Saeid Naderiparizi, Adam Scibior, Andreas Munk, Mehrdad Ghadiri, Atilim Gunes Baydin, Bradley J Gram-Hansen, Christian A Schroeder De Witt, Robert Zinkov, Philip Torr, Tom Rainforth, et al. Amortized rejection sampling in universal probabilistic programming. In International Conference on Artificial Intelligence and Statistics, pages 8392–8412. PMLR, 2022

    2022

  32. [32]

    Don’t be so negative! score-based generative modeling with oracle-assisted guidance

    Saeid Naderiparizi, Xiaoxuan Liang, Setareh Cohan, Berend Zwartsenberg, and Frank Wood. Don’t be so negative! score-based generative modeling with oracle-assisted guidance. In Forty-first International Conference on Machine Learning, 2024. 11

    2024

  33. [33]

    Constrained generative modeling with manually bridged diffusion models

    Saeid Naderiparizi, Xiaoxuan Liang, Berend Zwartsenberg, and Frank Wood. Constrained generative modeling with manually bridged diffusion models. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 19607–19615, 2025

    2025

  34. [34]

    A diffusion- model of joint interactive navigation.Advances in Neural Information Processing Systems, 36, 2024

    Matthew Niedoba, Jonathan Lavington, Yunpeng Liu, Vasileios Lioutas, Justice Sefas, Xiaoxuan Liang, Dylan Green, Setareh Dabiri, Berend Zwartsenberg, Adam Scibior, et al. A diffusion- model of joint interactive navigation.Advances in Neural Information Processing Systems, 36, 2024

    2024

  35. [35]

    Input perturbation reduces exposure bias in diffusion models.arXiv preprint arXiv:2301.11706, 2023

    Mang Ning, Enver Sangineto, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Input perturbation reduces exposure bias in diffusion models.arXiv preprint arXiv:2301.11706, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  37. [37]

    Constraining generative models for engineering design with negative data.Transactions on Machine Learning Research, 2024

    Lyle Regenwetter, Giorgio Giannone, Akash Srivastava, Dan Gutfreund, and Faez Ahmed. Constraining generative models for engineering design with negative data.Transactions on Machine Learning Research, 2024

    2024

  38. [38]

    Cambridge university press, 2000

    Leonard CG Rogers and David Williams.Diffusions, markov processes, and martingales: Volume 1, foundations. Cambridge university press, 2000

    2000

  39. [39]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  40. [40]

    Motion planning with sequential convex optimization and convex collision checking.The International Journal of Robotics Research, 33(9):1251–1270, 2014

    John Schulman, Yan Duan, Jonathan Ho, Alex Lee, Ibrahim Awwal, Henry Bradlow, Jia Pan, Sachin Patil, Ken Goldberg, and Pieter Abbeel. Motion planning with sequential convex optimization and convex collision checking.The International Journal of Robotics Research, 33(9):1251–1270, 2014

    2014

  41. [41]

    Understanding and improving training-free loss-based diffusion guidance.Advances in Neural Information Processing Systems, 37:108974–109002, 2024

    Yifei Shen, Xinyang Jiang, Yifan Yang, Yezhen Wang, Dongqi Han, and Dongsheng Li. Understanding and improving training-free loss-based diffusion guidance.Advances in Neural Information Processing Systems, 37:108974–109002, 2024

    2024

  42. [42]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. PMLR, 2015

    2015

  43. [43]

    Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    2019

  44. [44]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021

    2021

  45. [45]

    Attention is all you need.Advances in Neural Information Processing Systems, 2017

    A Vaswani. Attention is all you need.Advances in Neural Information Processing Systems, 2017

    2017

  46. [46]

    A connection between score matching and denoising autoencoders.Neural computation, 23(7):1661–1674, 2011

    Pascal Vincent. A connection between score matching and denoising autoencoders.Neural computation, 23(7):1661–1674, 2011

    2011

  47. [47]

    Diffusion-es: Gradient-free planning with diffusion for autonomous driving and zero-shot instruction following.arXiv preprint arXiv:2402.06559, 2024

    Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki. Diffusion-es: Gradient-free planning with diffusion for autonomous driving and zero-shot instruction following.arXiv preprint arXiv:2402.06559, 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Guidance with spherical gaussian constraint for conditional diffusion.arXiv preprint arXiv:2402.03201, 2024

    Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion.arXiv preprint arXiv:2402.03201, 2024

    Pith/arXiv arXiv 2024

  49. [49]

    Tfg: Unified training-free guidance for diffusion models.Advances in Neural Information Processing Systems, 37:22370–22417, 2024

    Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, and Stefano Ermon. Tfg: Unified training-free guidance for diffusion models.Advances in Neural Information Processing Systems, 37:22370–22417, 2024. 12

    2024

  50. [50]

    Freedom: Training- free energy-guided conditional diffusion model

    Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training- free energy-guided conditional diffusion model. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174–23184, 2023

    2023

  51. [51]

    Gradient surgery for multi-task learning.Advances in neural information processing systems, 33:5824–5836, 2020

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning.Advances in neural information processing systems, 33:5824–5836, 2020

    2020

  52. [52]

    Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps.arXiv preprint arXiv:1910.03088, 2019

    Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kummerle, Hendrik Konigshof, Christoph Stiller, Arnaud de La Fortelle, et al. Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps.arXiv preprint arXiv:1910.03088, 2019

    Pith/arXiv arXiv 1910

  53. [53]

    Motiondiffuse: Text-driven human motion generation with diffusion model.arXiv preprint arXiv:2208.15001, 2022

    Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model.arXiv preprint arXiv:2208.15001, 2022

    Pith/arXiv arXiv 2022

  54. [54]

    Diffusion-based planning for autonomous driving with flexible guidance.arXiv preprint arXiv:2501.15564, 2025

    Yinan Zheng, Ruiming Liang, Kexin Zheng, Jinliang Zheng, Liyuan Mao, Jianxiong Li, Weihao Gu, Rui Ai, Shengbo Eben Li, Xianyuan Zhan, et al. Diffusion-based planning for autonomous driving with flexible guidance.arXiv preprint arXiv:2501.15564, 2025. 13 A Algorithm Algorithm 1Rollout-Based Constrained Fine-Tuning Require: Pretrained denoiser D, fine-tuned...

    Pith/arXiv arXiv 2025